Adaptive action supervision in reinforcement learning from real-world multi-agent demonstrations